Spaces:
Sleeping
Sleeping
File size: 10,867 Bytes
33d4721 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 |
import os
import torch
from accelerate import PartialState
from huggingface_hub import HfApi
from peft import LoraConfig, PeftModel, get_peft_model, prepare_model_for_kbit_training
from transformers import AutoConfig, BitsAndBytesConfig, PaliGemmaForConditionalGeneration
from autotrain import logger
from autotrain.trainers.common import (
ALLOW_REMOTE_CODE,
LossLoggingCallback,
TrainStartCallback,
UploadLogs,
pause_space,
remove_autotrain_data,
save_training_params,
)
TARGET_MODULES = {}
SUPPORTED_MODELS = [
"PaliGemmaForConditionalGeneration",
# "Florence2ForConditionalGeneration", support later
]
MODEL_CARD = """
---
tags:
- autotrain
- text-generation-inference
- image-text-to-text
- text-generation{peft}
library_name: transformers{base_model}
license: other{dataset_tag}
---
# Model Trained Using AutoTrain
This model was trained using AutoTrain. For more information, please visit [AutoTrain](https://hf.co/docs/autotrain).
# Usage
```python
# you will need to adjust code if you didnt use peft
from PIL import Image
from transformers import PaliGemmaForConditionalGeneration, PaliGemmaProcessor
import torch
import requests
from peft import PeftModel
base_model_id = BASE_MODEL_ID
peft_model_id = THIS_MODEL_ID
max_new_tokens = 100
text = "Whats on the flower?"
img_url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/bee.JPG?download=true"
image = Image.open(requests.get(img_url, stream=True).raw)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
base_model = PaliGemmaForConditionalGeneration.from_pretrained(base_model_id)
processor = PaliGemmaProcessor.from_pretrained(base_model_id)
model = PeftModel.from_pretrained(base_model, peft_model_id)
model.merge_and_unload()
model = model.eval().to(device)
inputs = processor(text=text, images=image, return_tensors="pt").to(device)
with torch.inference_mode():
generated_ids = model.generate(
**inputs,
max_new_tokens=max_new_tokens,
do_sample=False,
)
result = processor.batch_decode(generated_ids, skip_special_tokens=True)
print(result)
```
"""
def get_target_modules(config):
if config.target_modules is None:
return TARGET_MODULES.get(config.model)
if config.target_modules.strip() == "":
return TARGET_MODULES.get(config.model)
if config.target_modules.strip().lower() == "all-linear":
return "all-linear"
return config.target_modules.split(",")
def create_model_card(config):
if config.peft:
peft = "\n- peft"
else:
peft = ""
if config.data_path == f"{config.project_name}/autotrain-data" or os.path.isdir(config.data_path):
dataset_tag = ""
else:
dataset_tag = f"\ndatasets:\n- {config.data_path}"
if os.path.isdir(config.model):
base_model = ""
else:
base_model = f"\nbase_model: {config.model}"
model_card = MODEL_CARD.format(
dataset_tag=dataset_tag,
peft=peft,
base_model=base_model,
)
return model_card.strip()
def check_model_support(config):
api = HfApi(token=config.token)
model_info = api.model_info(config.model)
architectures = model_info.config.get("architectures", [])
for arch in architectures:
if arch in SUPPORTED_MODELS:
return True
return False
def configure_logging_steps(config, train_data, valid_data):
logger.info("configuring logging steps")
if config.logging_steps == -1:
if config.valid_split is not None:
logging_steps = int(0.2 * len(valid_data) / config.batch_size)
else:
logging_steps = int(0.2 * len(train_data) / config.batch_size)
if logging_steps == 0:
logging_steps = 1
if logging_steps > 25:
logging_steps = 25
config.logging_steps = logging_steps
else:
logging_steps = config.logging_steps
logger.info(f"Logging steps: {logging_steps}")
return logging_steps
def configure_training_args(config, logging_steps):
logger.info("configuring training args")
training_args = dict(
output_dir=config.project_name,
per_device_train_batch_size=config.batch_size,
per_device_eval_batch_size=config.batch_size,
learning_rate=config.lr,
num_train_epochs=config.epochs,
eval_strategy=config.eval_strategy if config.valid_split is not None else "no",
logging_steps=logging_steps,
save_total_limit=config.save_total_limit,
save_strategy=config.eval_strategy if config.valid_split is not None else "no",
gradient_accumulation_steps=config.gradient_accumulation,
report_to=config.log,
auto_find_batch_size=config.auto_find_batch_size,
lr_scheduler_type=config.scheduler,
optim=config.optimizer,
warmup_ratio=config.warmup_ratio,
weight_decay=config.weight_decay,
max_grad_norm=config.max_grad_norm,
push_to_hub=False,
load_best_model_at_end=True if config.valid_split is not None else False,
ddp_find_unused_parameters=False,
gradient_checkpointing=not config.disable_gradient_checkpointing,
remove_unused_columns=False,
)
if not config.disable_gradient_checkpointing:
if config.peft and config.quantization in ("int4", "int8"):
training_args["gradient_checkpointing_kwargs"] = {"use_reentrant": True}
else:
training_args["gradient_checkpointing_kwargs"] = {"use_reentrant": False}
if config.mixed_precision == "fp16":
training_args["fp16"] = True
if config.mixed_precision == "bf16":
training_args["bf16"] = True
return training_args
def get_callbacks(config):
callbacks = [UploadLogs(config=config), LossLoggingCallback(), TrainStartCallback()]
return callbacks
def get_model(config):
logger.info("loading model config...")
model_config = AutoConfig.from_pretrained(
config.model,
token=config.token,
trust_remote_code=ALLOW_REMOTE_CODE,
use_cache=config.disable_gradient_checkpointing,
)
logger.info("loading model...")
if config.peft:
if config.quantization == "int4":
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=False,
)
elif config.quantization == "int8":
bnb_config = BitsAndBytesConfig(load_in_8bit=True)
else:
bnb_config = None
model = PaliGemmaForConditionalGeneration.from_pretrained(
config.model,
config=model_config,
token=config.token,
quantization_config=bnb_config,
trust_remote_code=ALLOW_REMOTE_CODE,
)
else:
model = PaliGemmaForConditionalGeneration.from_pretrained(
config.model,
config=model_config,
token=config.token,
trust_remote_code=ALLOW_REMOTE_CODE,
)
logger.info(f"model dtype: {model.dtype}")
if config.peft:
logger.info("preparing peft model...")
if config.quantization is not None:
gradient_checkpointing_kwargs = {}
if not config.disable_gradient_checkpointing:
if config.quantization in ("int4", "int8"):
gradient_checkpointing_kwargs = {"use_reentrant": True}
else:
gradient_checkpointing_kwargs = {"use_reentrant": False}
model = prepare_model_for_kbit_training(
model,
use_gradient_checkpointing=not config.disable_gradient_checkpointing,
gradient_checkpointing_kwargs=gradient_checkpointing_kwargs,
)
else:
model.enable_input_require_grads()
peft_config = LoraConfig(
r=config.lora_r,
lora_alpha=config.lora_alpha,
lora_dropout=config.lora_dropout,
bias="none",
task_type="CAUSAL_LM",
target_modules=get_target_modules(config),
)
model = get_peft_model(model, peft_config)
for param in model.vision_tower.parameters():
param.requires_grad = False
for param in model.multi_modal_projector.parameters():
param.requires_grad = False
return model
def merge_adapter(base_model_path, target_model_path, adapter_path):
logger.info("Loading adapter...")
model = PaliGemmaForConditionalGeneration.from_pretrained(
base_model_path,
torch_dtype=torch.float16,
low_cpu_mem_usage=True,
trust_remote_code=ALLOW_REMOTE_CODE,
)
model = PeftModel.from_pretrained(model, adapter_path)
model = model.merge_and_unload()
logger.info("Saving target model...")
model.save_pretrained(target_model_path)
def post_training_steps(config, trainer):
logger.info("Finished training, saving model...")
trainer.model.config.use_cache = True
trainer.save_model(config.project_name)
model_card = create_model_card(config)
# save model card to output directory as README.md
with open(f"{config.project_name}/README.md", "w", encoding="utf-8") as f:
f.write(model_card)
if config.peft and config.merge_adapter:
logger.info("Merging adapter weights...")
try:
del trainer
torch.cuda.empty_cache()
merge_adapter(
base_model_path=config.model,
target_model_path=config.project_name,
adapter_path=config.project_name,
)
# remove adapter weights: adapter_*
for file in os.listdir(config.project_name):
if file.startswith("adapter_"):
os.remove(f"{config.project_name}/{file}")
except Exception as e:
logger.warning(f"Failed to merge adapter weights: {e}")
logger.warning("Skipping adapter merge. Only adapter weights will be saved.")
if config.push_to_hub:
if PartialState().process_index == 0:
# remove data folder
remove_autotrain_data(config)
logger.info("Pushing model to hub...")
save_training_params(config)
api = HfApi(token=config.token)
api.create_repo(
repo_id=f"{config.username}/{config.project_name}", repo_type="model", private=True, exist_ok=True
)
api.upload_folder(
folder_path=config.project_name,
repo_id=f"{config.username}/{config.project_name}",
repo_type="model",
)
if PartialState().process_index == 0:
pause_space(config)
|